bool solve(const_Matrix<T, Block0> b, Matrix<T, Block1> x)
{
typedef typename Block_layout<Block0>::order_type order_type;
typedef typename Block_layout<Block0>::complex_type complex_type;
typedef Layout<2, order_type, Stride_unit_dense, complex_type> data_LP;
typedef Strided<2, T, data_LP, Local_map> block_type;
assert(b.size(0) == length_);
assert(b.size(0) == x.size(0) && b.size(1) == x.size(1));
Matrix<T, block_type> b_int(b.size(0), b.size(1));
assign_local(b_int, b);
if (tr == mat_conj ||
(tr == mat_trans && Is_complex<T>::value) ||
(tr == mat_herm && !Is_complex<T>::value))
VSIP_IMPL_THROW(unimplemented(
"LU solver (CVSIP backend) does not implement this transformation"));
{
Ext_data<block_type> b_ext(b_int.block());
cvsip::View<2,T,true>
cvsip_b_int(b_ext.data(),0,b_ext.stride(0),b_ext.size(0),
b_ext.stride(1),b_ext.size(1));
cvsip_b_int.block().admit(true);
traits::lu_solve(lu_, tr, cvsip_b_int.ptr());
cvsip_b_int.block().release(true);
}
assign_local(x, b_int);
return true;
}
static void add(
const_Matrix<TR, BlockR> res,
const_Matrix<T1, Block1> op1,
const_Matrix<T2, Block2> op2)
{
vsip::dda::Data<BlockR, vsip::dda::out> raw_res(res.block());
vsip::dda::Data<Block1, vsip::dda::in> raw1(op1.block());
vsip::dda::Data<Block2, vsip::dda::in> raw2(op2.block());
// int cost = raw_res.cost + raw1.cost + raw2.cost;
// cout << "Tag_plain " << cost << endl;
float *pR = raw_res.ptr();
float const *p1 = raw1.ptr();
float const *p2 = raw2.ptr();
for (index_type c=0; c<res.size(1); ++c)
{
for (index_type r=0; r<res.size(0); ++r)
{
pR[r*raw_res.stride(0) + c*raw_res.stride(1)] =
p1[r*raw1.stride(0) + c*raw1.stride(1)] +
p2[r*raw2.stride(0) + c*raw2.stride(1)];
}
}
}
static void add(
const_Matrix<TR, BlockR> res,
const_Matrix<T1, Block1> op1,
const_Matrix<T2, Block2> op2)
{
typedef typename BlockR::layout_type layout_type;
// Check that no memory is required.
// test_assert((dda::Data<BlockR, layout_type>::CT_Mem_not_req));
// test_assert((dda::Data<Block1, layout_type>::CT_Mem_not_req));
// test_assert((dda::Data<Block2, layout_type>::CT_Mem_not_req));
vsip::dda::Data<BlockR, vsip::dda::out, layout_type> raw_res(res.block());
vsip::dda::Data<Block1, vsip::dda::in, layout_type> raw1(op1.block());
vsip::dda::Data<Block2, vsip::dda::in, layout_type> raw2(op2.block());
// int cost = raw_res.cost + raw1.cost + raw2.cost;
// cout << "Tag_contig " << cost << endl;
float* pR = raw_res.ptr();
float* p1 = raw1.ptr();
float* p2 = raw2.ptr();
for (index_type i=0; i<res.size(); ++i)
{
*pR = *p1 + *p2;
++pR;
++p1;
++p2;
}
}
void
generic_prodj(
const_Matrix<T0, Block0> a,
const_Matrix<T1, Block1> b,
Matrix<T2, Block2> r)
{
assert(r.size(0) == a.size(0));
assert(r.size(1) == b.size(1));
assert(a.size(1) == b.size(0));
#ifdef VSIP_IMPL_REF_IMPL
impl::generic_prod(a, conj(b), r);
#else
vsip_csl::dispatch<vsip_csl::dispatcher::op::prod_mm_conj, void,
Block2&, Block0 const&, Block1 const&>
(r.block(), a.block(), b.block());
#endif
}
void
interpolate(
const_Matrix<IT, Block1> indices, // n x m
Tensor<T, Block2> window, // n x m x I
const_Matrix<complex<T>, Block3> in, // n x m
Matrix<complex<T>, Block4> out, // nx x m
length_type depth,
length_type padded_depth)
{
// All blocks must have the same dimension ordering
typedef typename Block_layout<Block1>::order_type order1_type;
typedef typename Block_layout<Block2>::order_type order2_type;
typedef typename Block_layout<Block3>::order_type order3_type;
typedef typename Block_layout<Block4>::order_type order4_type;
assert(order1_type::impl_dim0 == order2_type::impl_dim0);
assert(order1_type::impl_dim0 == order3_type::impl_dim0);
assert(order1_type::impl_dim0 == order4_type::impl_dim0);
assert(order1_type::impl_dim1 == order2_type::impl_dim1);
assert(order1_type::impl_dim1 == order3_type::impl_dim1);
assert(order1_type::impl_dim1 == order4_type::impl_dim1);
Device_memory<Block1> dev_indices(indices.block(), impl::SYNC_IN);
Device_memory<Block2> dev_window(window.block(), impl::SYNC_IN);
Device_memory<Block3> dev_in(in.block(), impl::SYNC_IN);
Device_memory<Block4> dev_out(out.block(), impl::SYNC_OUT);
size_t rows_in = in.size(0);
size_t rows_out = out.size(0);
size_t cols = in.size(1);
assert(cols == out.size(1));
interpolate(
dev_indices.data(),
dev_window.data(),
reinterpret_cast<cuComplex const*>(dev_in.data()),
reinterpret_cast<cuComplex*>(dev_out.data()),
depth,
padded_depth,
rows_in,
rows_out,
cols);
}
void
generic_prod(
const_Matrix<T0, Block0> a,
const_Vector<T1, Block1> b,
Vector<T2, Block2> r)
{
using namespace vsip_csl::dispatcher;
assert(r.size() == a.size(0));
assert(a.size(1) == b.size());
#ifdef VSIP_IMPL_REF_IMPL
Evaluator<op::prod_mv, be::cvsip,
void(Block2&, Block0 const&, Block1 const&)>::exec
(r.block(), a.block(), b.block());
#else
vsip_csl::dispatch<op::prod_mv, void,
Block2&, Block0 const&, Block1 const&>
(r.block(), a.block(), b.block());
#endif
}
void
matrix_add_1(
const_Matrix<TR, BlockR> res,
const_Matrix<T1, Block1> op1,
const_Matrix<T2, Block2> op2)
{
vsip::dda::Data<BlockR, vsip::dda::out> raw_res(res.block());
float *p_raw = raw_res.ptr();
stride_type row_str_raw = raw_res.stride(0);
stride_type col_str_raw = raw_res.stride(1);
vsip::dda::Data<Block1, vsip::dda::in> raw1(op1.block());
float const *p1 = raw1.ptr();
stride_type row_str1 = raw1.stride(0);
stride_type col_str1 = raw1.stride(1);
vsip::dda::Data<Block2, vsip::dda::in> raw2(op2.block());
float const *p2 = raw2.ptr();
stride_type row_str2 = raw2.stride(0);
stride_type col_str2 = raw2.stride(1);
for (index_type r=0; r<res.size(0); ++r)
{
float* row_raw = p_raw;
float const *row_1 = p1;
float const *row_2 = p2;
for (index_type c=0; c<res.size(1); ++c)
{
*row_raw = *row_1 + *row_2;
row_1 += col_str1;
row_2 += col_str2;
row_raw += col_str_raw;
}
p_raw += row_str_raw;
p1 += row_str1;
p2 += row_str2;
}
}
typename vsip::impl::scalar_of<T>::type
norm_1(const_Matrix<T, Block> m)
{
typedef typename vsip::impl::scalar_of<T>::type scalar_type;
scalar_type norm = sumval(mag(m.col(0)));
for (index_type j=1; j<m.size(1); ++j)
{
norm = std::max(norm, sumval(mag(m.col(j))));
}
return norm;
}
double
error_db(const_Matrix<T1, Block1> v1,
const_Matrix<T2, Block2> v2)
{
double maxsum = -250;
for (unsigned i = 0; i < v1.size(0); ++i)
{
double sum = error_db(v1.row(i), v2.row(i));
if (sum > maxsum)
maxsum = sum;
}
return maxsum;
}
inline bool equal(const_Matrix<T1, B1> v, const_Matrix<T2, B2> w)
{
if (v.size(0) != w.size(0) || v.size(1) != w.size(1)) return false;
for (length_type i = 0; i != v.size(0); ++i)
for (length_type j = 0; j != v.size(1); ++j)
if (!equal(v.get(i, j), w.get(i, j)))
return false;
return true;
}
void
matrix_add_2(
const_Matrix<TR, BlockR> res,
const_Matrix<T1, Block1> op1,
const_Matrix<T2, Block2> op2)
{
vsip::dda::Data<BlockR, vsip::dda::out> raw_res(res.block());
vsip::dda::Data<Block1, vsip::dda::in> raw1(op1.block());
vsip::dda::Data<Block2, vsip::dda::in> raw2(op2.block());
float *pR = raw_res.ptr();
float const *p1 = raw1.ptr();
float const *p2 = raw2.ptr();
for (index_type r=0; r<res.size(0); ++r)
{
for (index_type c=0; c<res.size(1); ++c)
{
pR[r*raw_res.stride(0) + c*raw_res.stride(1)] =
p1[r*raw1.stride(0) + c*raw1.stride(1)] +
p2[r*raw2.stride(0) + c*raw2.stride(1)];
}
}
}
void
interpolate(
const_Matrix<IT, Block1> indices, // n x m
Tensor<T, Block2> window, // n x m x I
const_Matrix<complex<T>, Block3> in, // n x m
Matrix<complex<T>, Block4> out, // nx x m
length_type depth)
{
length_type n = indices.size(0);
length_type m = indices.size(1);
length_type nx = out.size(0);
length_type I = depth; // window.size(2) may include padding
assert(n == in.size(0));
assert(m == in.size(1));
assert(m == out.size(1));
assert(window.size(0) == n);
assert(window.size(1) == m);
out = complex<T>(0);
for (index_type j = 0; j < m; ++j)
{
for (index_type i = 0; i < n; ++i)
{
index_type ikxrows = indices.get(i, j);
index_type i_shift = (i + n/2) % n;
for (index_type h = 0; h < I; ++h)
{
out.put(ikxrows + h, j, out.get(ikxrows + h, j) +
(in.get(i_shift, j) * window.get(i, j, h)));
}
}
out.col(j)(Domain<1>(j%2, 2, nx/2)) *= T(-1);
}
}